A08 Solutions

Reference answers with the reasoning. Try ../exercises/ first — the learning is in the struggle, not the answer key.

Nothing to compile; run them directly or from the module root:

python 01_fused_relu_add_solution.py
python 02_masked_mean_solution.py

cd .. && make run-solutions          # both, must all PASS

#	File	Notes
1	01_fused_relu_add_solution.py	tl.maximum(x + y, 0.0) between the loads and the store. Adds a tail case (n not a multiple of BLOCK_SIZE, sums exactly 0.0). Explains why the measured speedup is below the 40% traffic model.
2	02_masked_mean_solution.py	Bounds mask vs. data mask; other= as the reduction's identity; the empty row guarded by clamping the denominator, not just selecting the result. Adds a ragged-width case.
3	03_triton_vs_cuda_solution.md	Real line counts for the vector-add pair, a control-surrendered table with the one row that justifies keeping a C++ kernel, how to see each compiler decision, and a full review comment for a Triton rewrite that is 6% slower.

Both scripts must print PASSED on every check. If one does not on your machine, that is a curriculum bug worth reporting — not an exercise.